With an increasing use of digital cameras, along with the digitization of existing photograph collections, it is not uncommon for a personal image collection to contain many thousands of images.
The high number of images usually renders their classification in smaller sub-sets necessary to avoid a fastidious browsing. A classification can typically be based on textual annotation of the images, on labels or tags in the form of metadata added to images, or any other form of static classification requiring user explicit input. Such classifications however suffer from drawbacks. One drawback is that a classification corresponding to the taste of a given user of the collection of images, at the current moment in time, does not necessarily correspond to the taste of other possible users or the same user at another time, or different context of use. Another drawback is that many images/photographs may remain unlabelled or unclassified, since a classification requiring user input usually appears as a fastidious task.
To save the user from the classification efforts, some retrieval systems rely on objective image content analysis rather than existing image labels or tags. Data on image content likely to be used for classification or image retrieval may result from so called low level analysis or high level analysis of the image. A low level analysis comprises, for example, colours analysis, spatial frequency analysis, texture analysis, or histograms analysis, etc. High level analysis rather involves algorithms to derive information from the semantic content of the images. As an example of high level analysis, processing engines may be used to identify in an image semantic content such as human faces, skin, animals, sky, water, sea, grass etc.
The semantic content as well as some other low level features mentioned above may in turn be used separately or in combination to calculate a similarity between images and finally to classify the images based on their similarity.
As an illustration of existing classification and image searching methods, U.S. Publication No. 2006/0050933; U.S. Pat. No. 7,043,474 and U.S. Pat. No. 6,922,699 may be referred to.